The proposed model incorporates four publicly available datasets for four disease classes and one normal class, totaling 19,893 CT scans. Images from the SARS-CoV-2 CT dataset include 2482 Covid-19 classes and 1252 non-Covid-19 classes. The pneumonia images are drawn from the COVIDx-CT collection, which contains 4273 pneumonia images as well as 1583 normal images. The tuberculosis CT imaging dataset [28] used in this study includes 501 aberrant and 501 normal images obtained from 223 patients and manually graded by two professionals. Lung cancer images are collected from the cancer imaging archive (CIA) dataset. The dataset comprises a collection of lung CT images taken from the National Cancer Institute and connected with proteomic and genetic clinical data. The collection contains 5043 CT images obtained from 48 series. Table 1 provides a full summary of the dataset utilized in this research.

The image is encrypted in a medical IoT device and transferred to a fog node using the advanced encryption standard (AES) method [29]. AES is a symmetric block cipher that transforms a 128-bit image into ciphertext by adding a secret key with a length of 128, 256 or 192 bits. The number of encryption execution rounds varies depending on the size of the cipher key, such as 10, 14, or 12. The fog node decrypts the cipher image for further processing.

A fog node is utilized to connect the cloud nearer to the medical IoT devices at the network edge. In fog node cipher format image is decrypted and transformed into standard normalization by the given process:

Data augmentation [30]: Collected data volume is enhanced with a data augmentation technique to eliminate overfitting and construct a more generalized learning model. Augmentation methods such as flipping, rotation, cropping, reflection, scaling, and blurring with a Gaussian filter are applied to produce an augmented dataset

Resize: dataset images are varying in size; they are adjusted in the size of 224 × 224 and 229 × 229 pixels respectively to fit the chosen pre-trained models.

Noise removal [31]: To reduce additive noise from CT scans, a Wiener filter is used. Noises incorporated in the images during the acquisition process assist in false detection. Therefore, the noises must be removed to accurately diagnose lung diseases. To begin, a Wiener filter is employed to remove additive noise, where the gaussian noise is added with 0 mean and σ = 30 as a standard deviation, and the kernel size is set as 9 × 9. Followed by a median filter applied with a window size of 3 × 3 to remove salt and pepper noise from the image.

Transfer learning is an ML approach that uses a learned model from the source domain to boost generalization in a target domain. In this work Inception-ResNet-v2, ResNet-50, and VGG-19 pre-trained DL models are taken for fine-tuning. The model parameters are updated after each iteration to minimize the loss function in the training phase. The input size, parameters, and the layers detailed of the pre-trained models are shown in Table 2.

To train these models for reliable detection, a larger dataset is required, which will result in an appropriate global minimum for the cost function and it is derived as,(1) C=−1M∑i=1m⁡ln⁡(P(Yi|xi)) where the total number of training occurrences in the training set is represented by M. xj is the training set nth occurrence while Yi are its label and the probability of the classification is P(Yi|xi) .

In a sparser dataset, robust detection is more difficult to achieve, and the model is more prone to overfitting. So in transfer learning the initial layers weight parameters from the Inception-ResNet-v2, ResNet-50 and VGG-19 are reclaimed. The models are then fine-tuned using the acquired CT dataset by changing top layers and varied batch sizes, epochs, and learning rates. The pre-trained models’ primary layers contain general features, while the succeeding levels provide domain features. To make faster the learning of the newly added layers the learning rate of the top layers is assigned to a high value. Negative transfer does not occur in the proposed techniques due to the constrained size of the lung CT dataset. The proposed transfer learning architecture fully connected the final layer is set to 5 neurons to detect Cancer, pneumonia, tuberculosis, normal, and Covid-19. Fig. 2 depicts the fine-tuned pre-trained model. Generalizing the transfer learning model to prevent the overfitting purpose dropout 0.3 and 0.5 and batch normalization (BN) is added in the final dense layer. The softmax layer is chosen for binary classification, and its function is(2) σ(z→ )i = eZi∑j=1k⁡eZj

Here z→ is the vector input of the softmax function, the total number of classes is k and Zi depicts the vector elements. In the proposed model to detect the 4 types of diseases cause the learned parameters ( θs=θ1s,θ2s,θ3s, …, θns) is transferred from multi-source (mT≥1) for increasing the decision function fTjx (x = 1,…, mT ). The detail of the parameters used in the fine-tuning of each mode is given in Table 3. During the research VGG-19 and ResNet-50 are fine-tuned after 30 epochs, and Inception-ResNet-v2 is fine-tuned after 35 epochs. Each model’s loss function is calculated and optimized via the Adam optimizer and the hyperparameters are set as α = 0.001, β1=09, β2=0.999 and epsilon is set as e−8 .

Ensemble learning [32] can considerably improve a learning system’s predictive capabilities and reduce variance by combining numerous individual models into an intelligent ensemble. In this work, the selected heterogeneous DL models (Inception-ResNet-v2, VGG-19, ResNet-50) are combined through majority voting [33], weighted averaging, and stacking to classify lung diseases. Fig. 4 depicts the proposed ensemble classifier for 5 class classification problems. In majority voting, multiple transfer learned models’ predictions are assigned as votes. Individual classifiers vote for a specific class, and the class with the most votes determine the ensemble’s output. The model’s output label is ht(x) , and weights are wt . The set value is t =  {1, …, K} , the majority voting combiner outcome is(3) H(x)=sign(∑t=1K⁡wtht(x))

In weighted averaging the base learner’s prediction outputs are averaged with dissimilar weights as follows:(4) Y(x)=∑i=1N⁡wiy^i(X),    wi≥0 and ∑i=1N⁡wi=1

Herein Y(x) is the integrated forecasting output, and wi depicts the weight of the base learners, y^i(x) represents the ith point base learners forecasting output and the total number of base learners is N.

Stacking (stacked generalization) [34] generates the output by using a meta-learner [35], that combines the individual weak learner’s prediction. Stacking uses the base learner prediction as a context and then differentially weighs these predictions to perform better than the individual base learner.

The operation of the stacking is illustrated in Fig. 3.

The severity level of the classified result H(x) is detected by using RNN-LSTM. Severity detection is vital to diagnose the disease for taking appropriate treatment. The diseases Covid-19, TB, pneumonia, and lung cancer severity score levels depend on the level of disease spread over left and right lungs, and the different output scores are 1:1–25% (low), score 2:26–50% (mild), score 3:51–75% (severe), score 4: > 75% (more severe). RNN is a subset of ANN, which has a feedback loop called a recurrent loop that can process the sequence of data. RNN is differentiated by their memory, which allows them to impact contemporary input and output with the information it gathered from previous inputs. In the RNN-LSTM structure, the nonlinear function (tanh) of the RNN is replaced with LSTM which is a neuronal mechanism that performs a memory cell. The LSTM is linked to the RNN to solve the vanishing gradient problem that affects the RNN’s long-term temporal reliance. The structure of the RNN is shown in Fig. 5. RNN has 3 layers respectively Input, Recurrent hidden layer, and output layer. RNN-LSTM receives the input picture H(x) as a sequence of vectors with the time T, its weight matrix is GIH . The hidden layers of RNN are attached recurrent manner which is described as Wh={W1, W2, …, Wk} , k-a total hidden unit of the layer. The hidden layer system memory is represented as,(5) Wh= η(Jh) (6) Jh=GWIHμH(x)+GWHHμH(x)+∅

Jh is the representation of state-space or system memory, η depicts the activation function, bias of the hidden node is indicated as ∅ , The weight amid the input and hidden layers is denoted by WIH , WHH indicate the weight amid the hidden and hidden layers. Consequently, the final output vector is calculated as(7) Ox= η(WHo+b)

Herein Ox indicate the resultant vector, b specify the bias of the final layer. LSTM function in RNN is depicted as:(8) (hn, cn)=lstm(cn, hn−1, cn−1)

RNN generated scores represent the severity level which depends on the disease spread over the left and right lung. The target score is calculated with the softmax function. softmax_cross_entropy_with_logits_v2 predicts the loss function and an SGD optimizer is used to minimize the loss. Initially, the batch size is set as 8, the training epoch of the net is 20 and the learning rate is 10−4 .

The proposed framework is assessed using 5-fold cross-validation. In this dataset, 80% of the data is used for training, and 20% for testing. Performance of the suggested method is measured using the following metrics:

(9) Accuracy=TP+TNTP+TN+FP+FN

(10) Specificity=TNTN+FP

(11) Sensitivity=TPFN+TP

(12) F1 score=2×TP(2×TP+FP+FN)

Here True Positive (TP) is the correctly predicted lung diseases, False Negative (FN) is the casing in which the network predicts the lung disease incorrectly, True Negative (TN) is the correctly predicted non-diseases images, and False Positive (FP) are cases in which the network predicts the normal images as a disease. In the proposed framework MATLAB (2020b) is used for implementation. With 64 GB of RAM, the server system is powered by a 3.5 GHz Intel Xeon F8 CPU.

The 3 transfer learning models that are independently trained and attained accuracies for binary classification are VGG-19 (95.60%), ResNet-50 (94.4%), and Inception-ResNet-v2 (95.23%). For the proposed ETL architecture, the training time is estimated to be around 15 min and it achieved higher accuracy of 98.6% compared to the individual model. Stacking, voting, and averaging ensemble concepts are adopted for the proposed multiclass classification problem. The proposed ensemble transfer learning framework achieved 0.978 precision, 0.982 recalls, and 0.974 F1-score for predicting 5 class lung diseases, indicating that the proposed approach generalization ability is greater than existing models. The results reveal that the ETL framework achieves the greatest accuracy, implying that there is a low risk of an incorrect prediction. Table 4 contains the results of the individual learning models’ performance.

Table 5 shows the severity prediction framework RNN-LSTM model performance with specificity, sensitivity, F1 score, and accuracy, RNN model training and testing losses. The resulting probability score of the RNN-LSTM can then be evaluated with a threshold value to decide the predicted images belong to which stage of severity. The RNN-LSTM achieved 98% of severity detection. The specificity (0.97) of the RNN-LSTM showed the possibility of false prediction is significantly low so this method is effectively suitable for the severity analysis.

Fig. 6 depicts the proposed RNN-LSTM model’s training and testing loss curves. The x-axis shows training epochs ranging from 0 to 100, while the y-axis illustrates cross-entropy. The curve indicates that as batch size increases, the model’s loss function decreases.

The findings show that a proposed DL methodology may considerably aid in the identification of lung disease subjects using CT scan images, providing a low-cost and quick option for Covid-19 identification. Fig. 7 shows the accuracy of the individual learning networks for 0–100 epochs. The proposed model’s accuracy is substantially higher than other models, and it grew as the epoch size increased, as seen in the plot.

The confusion matrix plot for the individual pre-trained model and proposed framework is illustrated in Fig. 8. Confusion matrixes visualize and describe the efficiency of a classification model. Compared to the individual models, the proposed model performed best with an overall accuracy of 98.6% and a less misclassification rate of 0.0140 for five classes. Fig. 9 depicts the receiver operating characteristic (ROC) curve for the proposed model. The ROC plot is shown vs. the true positive rate (vertical axis) and false positive rate (horizontal axis). The proposed model had a high area under the ROC curve (AUC) value, proving its capability to distinguish between classes.

Table 6 compares the proposed multi-classification technique to existing state strategies in terms of accuracy. Although the notion in [32] closely follows the proposed technique, it only classifies two classes, which is lower than the ETL framework, and it does not provide severity analysis. Because of its low generality, autoencoder-based CNN [5] performs poorly. The accuracy of the [19,24] approaches is 0.09%, 1.1% lower than that of the proposed model. As a result, with a final classification rate of 98.6 percent, the proposed ensemble transfer learning for the multi-classification job surpassed numerous previous techniques.

Table 7 depicts the performance evaluation of the RNN-LSTM with existing state of art models. Reference [10] used ResNet-50 and DenseNet-201 hybrid model transfer learning method which is suitable for fewer datasets but training is complex also performance is lacking by 0.16% more than the proposed framework. Reference [17] used the adaptive fractional crow-oriented CNN model to figure out lung micro bacterial disease but the training time is high compared to the RNN-LSTM framework. The proposed model outperforms the tri-level attention based SegNet model [1] by 7.3%. The proposed RNN-LSTM model offers high accuracy, specificity, and F1 score compared to the existing state-of-art severity framework; these results prove to outline the use of its severity detection in lung diseases.